The AI sorts the inbox. Your team checks and sends.

In a shared business mailbox, AI does three jobs. It sorts each new email into a category with an owner, it drafts a reply or a quote, and it fills in the customer data in the CRM. A person checks the draft and sends it. The value comes from that order: the routine work is prepared, and the decision stays with your team.

Sorting works when categories, responsible teams and permitted follow-up actions are defined in advance. Clear rules come first: a dedicated mailbox or a valid case number decides before any model does. The AI helps with free text, varied wording and mixed requests. Its output stays small: a proposed category, a detected reference, a short reason and its uncertainty. Our guide to sorting an email inbox with AI shows a routing matrix and a worked example.

Start with a few categories such as invoice, customer enquiry, existing case, HR and unresolved. Each one needs an owner, the references it requires and the actions it permits. Keep priority separate from category: "urgent" in a subject line is a signal, not a deadline.

A conversation can contain several topics, so each message is recorded with its link to the business case. Attachments need their own check, because an email headed "Invoice" can contain an invoice and a contract. Our guides on classifying documents and saving Outlook attachments to SharePoint cover that step.

The mailbox connection decides which folders are watched and which changes start processing. Microsoft documents per-folder delta queries for message changes in Microsoft Graph, and its Outlook connector reference lists possible duplicate triggers and limits with protected messages. A repeated event must therefore not create a second task. Our articles on n8n with Outlook and Microsoft 365 and n8n with Google Workspace describe both mailbox types.

Customer data goes into the CRM as a contact, a case or a task. Existing contacts are matched first, so that the same customer is not created twice. Our guide to CRM automation without duplicate contacts shows the matching rules.

A reply draft is prepared work, not a sent email. By default, the mailbox connection gets no right to send, and a person approves recipient and content. An email that tells the system to ignore its rules is input data. It does not change which recipients and actions are permitted. Our article on draft plus sign-off explains who approves what, and requesting missing client documents shows a follow-up request with staff approval.

At Statzinger Steuerberatung, an online tax firm in Vienna, AI sorts incoming client emails into the firm's working categories and writes reply drafts at the press of a button, inside Outlook, BMD and the CRM. The firm always decides and sends. The triage ran in observation mode first and was released only after an acceptance test on at least 150 real emails, with a separate threshold for each type of error. The case study describes the criteria.

1. Pick one shared mailbox and a few explicit tasks.
2. Write down the categories, owners and permitted actions.
3. Collect real emails, including short replies, long threads, mixed requests and manipulated instructions.
4. Run the sorting in observation only and compare it with your team's labels.
5. Go live with a pause control and an owned queue for unclear mail.

Our AI email inbox page describes how a first project with us runs.